Qualitative research is changing, and most of the conversation about it is getting the direction wrong.
The worry you hear most often is that AI will devalue qualitative expertise. If a language model can read a thousand interviews and produce a summary, what exactly is the researcher for? This framing treats AI as a replacement for humans, and leads people to point to AI flaws to justify continued need for the human analyst. While AI cannot replace human judgement, the continuous advances in technology make this a risky approach. But the whole line of argumentation is missing the bigger point.
I believe AI is not something replacing human judgement, but rather a transformation of the conditions under which that judgement operates. To understand what is really happening, it helps to look at what happened at the quantitative side when computers arrived.
The quantitative precedent
In the 1940s, electronic computers began to take over the basic calculations that had previously required teams of human "computers" — people whose job was literally to perform arithmetic. By the 1980s, spreadsheets had arrived on every desktop, and the need for manual calculation in business and academia was essentially eliminated. What happened to quantitative expertise?
It did not disappear, it exploded. The models became far more complex. Businesses that had previously managed with simple cash flow projections began running discounted cash flow analyses, sensitivity tables, Monte Carlo simulations, and portfolio optimisation models. Academic economics produced new empirical methods that would have been computationally impossible without software. Finance invented derivatives pricing. Epidemiology developed survival analysis. The elimination of manual calculation did not reduce the demand for quantitative expertise — it increased it, because the tools made sophistication affordable.
When computers were introduced to do quantitative tasks, it did not go without resistance. As pocket calculators became affordable in the 1970s and spreadsheets followed in the early 1980s, a genuine public debate erupted over whether machines should replace human arithmetic. The calculator evoked fears that students' computational abilities would be ruined, that people would become too reliant on machines, and that they wouldn't learn from their errors. Accountants and auditors had genuine professional objections: while manual calculation created an audit trail that was understood step-by-step, early spreadsheet errors were hard to trace.
Ultimately of course the tools evolved and the computers became integral to quantitative analysis. The people who thrived were not those who had been fastest at arithmetic. They were those who understood the underlying models well enough to use the new tools intelligently and interpret the outputs critically. The people who suffered were those who confused fluency with numbers for genuine analytical skill.
The qualitative bottleneck
Qualitative research covers an enormous range. At one end: a researcher spending a year embedded in a community on a remote island, conducting ethnographic fieldwork, producing a deeply personal monograph that takes another five years to write. At the other end: a product manager who has a list with 20 customer feedback open text quotes and needs to identify what to fix next quarter. In between: clinical interviews, policy consultations, focus group interviews, journalistic investigations, strategy interviews, employee surveys, and countless other forms of inquiry that produce meaning rather than measurements.
What these methods share is a fundamental dependence on human reading and basic processing time. Unlike numbers, text could not be processed by computers in any meaningful sense. Early attempts at "text analysis" were essentially sophisticated word counting — frequency tables, keyword searches, concordances. These captured surface patterns but missed the thing that makes qualitative data valuable: meaning in context. A keyword analysis of interviews about healthcare reform will tell you that people mentioned "waiting times" frequently. It will not tell you whether they were resigned, outraged, or had developed practical workarounds, or whether the waiting times were a symptom of a deeper organisational problem they could articulate clearly but only if asked the right way.
Because human reading was the bottleneck, qualitative research scaled poorly. The methods that produced the richest insight — the year of fieldwork, the painstaking interview study with fifty participants — were also the most time-consuming. The constraint was not imagination or curiosity but hours in the day. Research that would have been immensely valuable was never done because nobody had the time to analyse the data even if they could collect it.
What changes with LLMs
Large language models do something that previous software could not: they process meaning. Not perfectly, not without limitations, not in ways that should be trusted uncritically (the hallucinations and context window constraints of basic chatbots are real), and not in the way humans do. But the fundamental capability is genuinely new: a language model going through an interview transcript is doing something categorically different from keyword counting. It is parsing sentences, inferring intent, recognising context, and connecting ideas across paragraphs in ways that resemble — imperfectly but recognisably — the cognitive work a human reader does.
This breaks the bottleneck. A researcher who could previously read and code fifty interviews in six weeks can now work with five hundred. An organisation that collected customer feedback from a few dozen annual interviews can collect it continuously from thousands of interactions. A policy team that sampled 10% of consultation responses because reading all of them was impossible can now process all of them, systematically, with documented methodology.
The question is what happens when the bottleneck breaks.
The renaissance
The history of transformative tools in research suggests a consistent pattern. When a binding constraint is removed, the field does not simply do the same work faster. It discovers entirely new questions that the constraint had made unthinkable. This is similar to what Jevons famously discovered in 1865 with coal - the more efficient the engines became the more coal was used in aggregate.
Epidemiology did not just run its existing studies more quickly after statistical computing arrived — it developed new longitudinal study designs and population-scale analyses that could not have existed before. Genomics did not just sequence existing target genes faster — it developed whole-genome sequencing and gave rise to an entirely new understanding of human disease. The removal of a bottleneck does not compress the existing field; it expands it.
Qualitative research is at the beginning of the same expansion. The methods that will develop from here will look different from what exists today. The always-on research model — continuous qualitative signal from customer interactions, patient conversations, or employee feedback — is already becoming practical. The combination of ethnographic depth with statistical breadth, previously a methodological contradiction, becomes coherent when the analysis burden is shared with AI. Research questions that required ten years and a team of graduate students to answer might take two years and one researcher and thus become feasible for new topics. Questions that required two years might take two months. And some questions that nobody thought to ask — because the answer seemed permanently out of reach — will start getting answers.
None of this happens automatically. The questions still have to be formulated by humans who understand what they are asking and why. The data still has to be collected in ways that are rigorous enough to support the conclusions researchers want to draw. And the outputs still have to be interpreted by someone with the domain knowledge to distinguish a genuine finding from an artefact of how the data was collected or the model was prompted.
Most critically, the field needs to ensure the new tools are harnessed in the right way.
There are some beliefs that need to be re-examined, for example related to sample sizes. The common argument is that since qualitative research aims to deeply understand meaning, not to generalise findings to a population, smaller samples are preferred. Studies focusing on a single case are best for building theories, while comparative case studies extend and test theory. This received wisdom is based on the assumption that there is trade-off between depth and breadth... what could be unlocked if that bottleneck was removed and one could have both? Could larger samples allow discovering emerging patterns and hearing also minority voices, or help surface differences between groups of people in new ways?
Equally, the quality bar needs to be kept high. While off-the-shelf AI tools like ChatGPT can produce text that looks like qualitative analysis, the lack of rigour and depth means they are not justifiable. Synthetic respondents can not be blindly used everywhere and their limitations must be clearly understood. AI-assisted interviewing is not the same as human-led interviews and should not be passed on as such. There is room for people who take a nuanced and clear-headed view towards the new tools and help develop methods that result in better outcomes, not speeding up the creation of slop.
Why expertise matters more
The chainsaw is a useful reference point here. Before the chainsaw, logging required physical strength, endurance, and a specific manual skill. The chainsaw eliminated those requirements but created new ones. An experienced logger with a chainsaw can do in a day what previously took a week. An inexperienced person with a chainsaw is not slightly less effective than an experienced one... they are a serious danger to themselves and others. The tool amplifies competence and incompetence in equal measure. And let's not even talk about forestry harvesters... letting amateurs like me operate them would be a recipe for disaster.
The same principle applies here. A skilled qualitative researcher using AI-assisted analysis can do work that would previously have been impossible. A researcher who does not understand the methodological principles and the domain being studied (e.g. what makes a good research question, how interview design shapes the data that comes back, what counts as genuine evidence versus plausible-sounding output) will produce results that look impressive and are wrong in ways that may not be immediately obvious.
The checklist for AI-assisted qualitative analysis published here is a practical attempt to codify the things an expert checks that a novice might miss. But a checklist is only useful to someone who already understands why each item matters. The underlying expertise — in thematic analysis methodology, in research design, in the epistemological questions that determine what kind of conclusions qualitative data can support — is more valuable in an AI-assisted environment, not less, because the consequences of getting it wrong are now more severe.
The skills that become less valuable are the ones that were always bottlenecks rather than genuine expertise: reading speed, the ability to hold large amounts of text in memory while coding manually, the patience to work through transcripts that are mostly repetitive. The skills that become more valuable are the ones that were always the point: knowing what to ask, knowing what counts as an answer, and knowing what the data cannot tell you.
This is not a comfortable position for everyone. People who built careers on the bottleneck skills will find the ground shifting. Some might turn to arguing why firewood made with the good old artisanal methods burns better. But for researchers who have always cared more about the questions than the coding, this is a genuinely exciting moment — the moment when the tools finally caught up with the ambition.
What this looks like in practice
The research designs that will be possible in ten years will be substantially different from those that are normal today. Some of these are already visible. AI-led interviews at scale make it possible to collect structured qualitative data from hundreds or thousands of participants without the cost and coordination overhead of human-led interviews — opening up research questions that previously required either massive grant funding or methodological compromise. Always-on qualitative research makes it possible to monitor customer experience or employee sentiment not as periodic snapshots but as a continuous signal, with the depth that interview data provides rather than the shallowness of a satisfaction score.
What is less visible but equally important is the methodological innovation that will follow. Researchers are creative. Once they understand what is now technically feasible, they will design studies that take advantage of it. The ethnographer who previously had to choose between depth and scale will find new hybrid approaches. The policy analyst who previously had to sample because reading everything was impossible will develop new frameworks for working with complete data. The clinical researcher who previously could only follow fifty patients in depth will start asking what becomes possible with five hundred.
This is what a renaissance looks like. Not everything at once, and not without real risks — the concern about AI-generated slop passing for insight is legitimate and the field will have to develop new norms for what counts as rigorous work. But the direction is unmistakable. The constraint is lifting. The questions that were previously unaskable are coming into reach.
We are entering a new era for qualitative research. I am personally excited to see the great works that will be produced in the coming decades!
About the author
Olli Salo is a former Partner at McKinsey & Company where he spent 18 years helping clients understand the markets and themselves, develop winning strategies and improve their operating models. He has done over 1000 client interviews and published over 10 articles on McKinsey.com and beyond. LinkedIn profile